Support NeoGAF

[ChiefDada]

You're putting a lot of pressure on that upscaling.

True, I am. Could be wrong and it's not as potent as I predict.

ChiefDada @winjer What do you think about the rumored 2.18GHz? I don't think I buy it honestly.

It's weird. Maybe it's similar to early PS5 dev kit that had 2ghz base clock profile? Surely they would stick with SmartShift again right?

 

[Gaiff]

True, I am. Could be wrong and it's not as potent as I predict.



It's weird. Maybe it's similar to early PS5 dev kit that had 2ghz base clock profile? Surely they would stick with SmartShift again right?

Yeah, I think it was the Gonzalo or Oberon. Forgot the name but this was when the 9 TFLOPs rumors started. The CU count was correct but the clocks were way lower.

Also, it seems that 60 CUs might not be true after all. It might be 60 CUs with 6 disabled for 54 CUs at 2.42GHz which would make far more sense. Cerny did prefer higher clocks so it'd be weird for him to go the opposite direction with the Pro.

 

[Topher]

Was this confirmed? Well, not "confirmed" but I don't recall seeing the CU count in the leaks.

I didn't see the CU count in the leaks.

 

[Quantum253]

I didn't see the CU count in the leaks.

I think around the 12 min mark he mentions 56CU @2.2ghz (10.3TFLOPs to 33.5 FP32). I think that's how he gets the 42% increase

 

[Proelite]

I like your numbers as thats the range I hinted at for quite awhile now as this is the numbers I was getting

You nailed it tbh. That's exactly what it is in rdna2 terms.

 

[Topher]

If we take the 7800 XT and the 7700 XT as reference:

60 CUs @ 2.18 Ghz = 33.5 TF
54 CUs @ 2.42 Ghz = 33.5 TF

Pretty sure you use the shader units to calculate TF

PS5, for example.

2304 shader units times 2 operations per cycle times 2233mhz (freq) = 2304x2x2233=10,289,000/1,000,000=10.28TF

AMD Playstation 5 GPU Specs

AMD Oberon, 2233 MHz, 2304 Cores, 144 TMUs, 64 ROPs, 16384 MB GDDR6, 1750 MHz, 256 bit

www.techpowerup.com

If we do the same calculation for the RDNA 2.0 6800 XT....

4608 shader units times 2 operations per cycle times 2250mhz = 4608x2x2250=20,736,000/1,000,000=20.73 TF

Gaiff what is interesting about this is if we do the same calculation on 7700 XT then we do not get 35 TF

3456 shader units times 2 operations per cycle times 2544mhz = 3456x2x2544 = 17,584,128/1,000,000=17.58 TF

So the only variable that can be questioned here is the operations per cycle which relates to the dual-issue shaders you were talking about. What is differnt in RDNA 3.0 vs 2.0. Techpowerup calculates it differently as well for RDNA 3.0

AMD Radeon RX 7700 XT Specs

AMD Navi 32, 2544 MHz, 3456 Cores, 216 TMUs, 96 ROPs, 12288 MB GDDR6, 2250 MHz, 192 bit

www.techpowerup.com

 

[Topher]

I think around the 12 min mark he mentions 56CU @2.2ghz (10.3TFLOPs to 33.5 FP32). I think that's how he gets the 42% increase

Hmm, he was trying to work the math backwards. Not sure it works out that way, but I was actually trying to find something in the leak doc that showed the CUs and did't see it.

Thanks for pointing that out though.

 

[Gaiff]

Pretty sure you use the shader units to calculate TF

There are 64 shaders per CUs so you'll get the same result. Just need to do CUs X 64.

If we do the same calculation for the RDNA 2.0 6800 XT....

4608 shader units times 2 operations per cycle times 2250mhz = 4608x2x2250=20,736,000/1,000,000=20.73 TF

Gaiff what is interesting about this is if we do the same calculation on 7700 XT then we do not get 35 TF

3456 shader units times 2 operations per cycle times 2544mhz = 3456x2x2544 = 17,584,128/1,000,000=17.58 TF

So the only variable that can be questioned here is the operations per cycle which relates to the dual-issue shaders you were talking about. What is differnt in RDNA 3.0 vs 2.0. Techpowerup calculates it differently as well for RDNA 3.0

AMD Radeon RX 7700 XT Specs

AMD Navi 32, 2544 MHz, 3456 Cores, 216 TMUs, 96 ROPs, 12288 MB GDDR6, 2250 MHz, 192 bit

www.techpowerup.com

Yeah, that's what I arrived at. Same performance as the 6800/7700 XT.

 

[Topher]

There are 64 shaders per CUs so you'll get the same result. Just need to do CUs X 64.

Ah....that makes sense now.

 

[Jesb]

Not sure if it’s been said. But it’s about fucken time.

 

[playsaves3]

nope

I swear I saw someone with your profile pic there

 

[Audiophile]

GDC upon us and leaks are coming. Hopefully this prompts Sony to make an announcement sooner rather than later..

 

[NeonGhost]

GDC upon us and leaks are coming. Hopefully this prompts Sony to make an announcement sooner rather than later..

They won’t

 

[farmerboy]

See, this shit is way above my paygrade

Thank you guys, I feel better, need to erase my wtf text I was sending

Instead, ask if they know anything about a Bloodborne remaster/remake by Bluepoint that Sony will use to market the PS5pro?

 

[playsaves3]

Not confident at all because RDNA2's SMIDs don't have dual-issue compute capabilities so it would presumably require a lot more effort from the developer to take advantage of it (so no benefit for the base PS5). As far as I'm aware, not a single PC game takes advantage of that. It could be a possibility in first-party games from Sony but again, would require a lot more work.

The expectation is that it's going to be 4x as fast as the regular PS5 in RT but that's assuming the part "have seen speed up to 4x" is feasible and not just some absolute maximum performance. If we get 4x the performance, awesome, but it's better to be conservative and assume 2x on the low-end or 3x on the high-end. What I'm saying is, let's be reasonable and conservative with our estimates rather than running with high-end numbers as if that's what we're going to get for sure.

I think it makes sense to think about 3x on average with 2x and 4x being outlier cases but I see what your saying

 

[TheRedRiders]

 

[Moses85]

For the dummies here. How many duct taped gamecubes are this?

 

[farmerboy]

For the dummies here. How many duct taped gamecubes are this?

 

[MisterXDTV]

GDC upon us and leaks are coming. Hopefully this prompts Sony to make an announcement sooner rather than later..

They want to still sell consoles from now until November....

Only Microsoft was so stupid to announce a console 2 years in advance.....

 

[MisterXDTV]

Actually Henderson never confirmed that the 60 CUs are all active

He said the PS5 Pro APU GPU has 60 CUs, there is a difference

He never mentioned 64 CUs either

 

[angrod14]

I’m so tired of the whole “AI” fad now. Everything’s AI this, AI that, even hardware upgrades. I want the true shit. Be real, bitch.

Seriously thinking about buying a typetwriter.

 

[Gaiff]

I’m so tired of the whole “AI” fad now. Everything’s AI this, AI that, even hardware upgrades. I want the true shit. Be real, bitch.

Seriously thinking about buying a typetwriter.

Lol if you think AI is a fad.

 

[MisterXDTV]

I’m so tired of the whole “AI” fad now. Everything’s AI this, AI that, even hardware upgrades. I want the true shit. Be real, bitch.

Seriously thinking about buying a typetwriter.

I guess we will go full circle: the world will come back to analog someday

 

[shamoomoo]

Actually Henderson never confirmed that the 60 CUs are all active

He said the PS5 Pro APU GPU has 60 CUs, there is a difference

He never mentioned 64 CUs either

It was most likely implied because recent consoles release hasn't offered a fully enabled GPU.

 

[mrBarrelNut]

Are you talking about in terms of pure rasterization? Seems like you could arrive at the number if you equated the PS5 to an RTX 2060 super then added 45% to arrive at 2080Ti levels, but in terms of TOPS, looks like the RTX2080Ti only has 114 Vs. 300 for the PS5 Pro (If specs are true).

What's this shit? Since when have we been using TOPS as an a benchmark for gaming performance? Anyways, which TOPS are you talking about, INT4 or INT8? I'm pretty sure AMD will inflate these numbers as high as they can so they'll use INT4 numbers.

This is a slide from 2018 or whenever Turing was released:

So 228 TOPS INT8 and 455 TOPS INT4 for RTX2080 Ti.

 

[FireFly]

It's weird. Maybe it's similar to early PS5 dev kit that had 2ghz base clock profile? Surely they would stick with SmartShift again right?

The 7800 XT only hit a 2425 MHz average at 250W of power consumption. It's hardly crazy to think that ~180W part would see reduced clock speeds. But people in these threads continue to believe insane clock speeds that aren't even seen on PC.

 

[sendit]

I’m so tired of the whole “AI” fad now. Everything’s AI this, AI that, even hardware upgrades. I want the true shit. Be real, bitch.

Seriously thinking about buying a typetwriter.

This individual thinks AI is a fad. I guess if you think the invention of the calculator was a fad.

 

[MasterCornholio]

Which brings us back to that 9tf number I mentioned. Pretty big deal was made about that at the time when compared to XSX's 12tf.

Yep the system was called weak station 5 back then by the shills. Unfortunately for them it turned out much better than that. They have been suffering ever since.



Anyways I'm just happy that both systems turned out great.

 

[angrod14]

This individual thinks AI is a fad. I guess if you think the invention of the calculator was a fad.

Calculators don’t try to pass themselves as something they’re not.

 

[Topher]

Calculators don’t try to pass themselves as something they’re not.

How is AI being misrepresented here?

 

[Killjoy-NL]

Think you're wrong that the situation is much different. Just because they're happy to compromise first party exclusivity by releasing on PC, doesn't mean they'd be happy to lose PS customers to PC so long as they still buy the games on Steam. They still want to avoid that scenario because they lose out on all the commission.

In fact, there's even more imperative now to release a Pro because with their games being on the PC too there's even more temptation for those hardcore gamers to switch.

So the objective is in fact the same: stop the mid gen flow of PS gamers to PC.

It's a dumb reason though. If someone really wants to jump to PC, a Pro-console would be too little of a jump to warrant a release for that specific reason.

 

[sendit]

Calculators don’t try to pass themselves as something they’re not.

Humans can calculate math without calculators. Calculators are doing defined human things, math. Your argument is flawed here.

 

[angrod14]

How is AI being misrepresented here?

Native content (real deal) vs. upscaled chit.

 

[angrod14]

Humans can calculate math without calculators. Calculators are doing defined human things, math. Your argument is flawed here.

Pure mathematical tasks. No complex reasoning or creative output.

 

[Topher]

Native content (real deal) vs. upscaled chit.

When was the last time you played a AAA game on console that did not incorporate upscaling?

 

[rickybooby87]

What's this shit? Since when have we been using TOPS as an a benchmark for gaming performance? Anyways, which TOPS are you talking about, INT4 or INT8? I'm pretty sure AMD will inflate these numbers as high as they can so they'll use INT4 numbers.

This is a slide from 2018 or whenever Turing was released:

So 228 TOPS INT8 and 455 TOPS INT4 for RTX2080 Ti.

So the TOPS stats are incorrect for the RTX 2080Ti that you are referring to. the below screen shot is direct from NVIDIA:

NVIDIA GeForce RTX 4060 Ti & 4060 Graphics Cards

New ways to create & much more.

www.nvidia.com

(Once on the above link, go to specs then you can compare against older cards).


Additionally, from the PS5 Pro leaks, it's INT8 as specified in the documentation. For techniques like DLSS, tensor cores/TOPS is king. Sony basically has hardware accelerated upscaling/reconstruction support now.

 

[angrod14]

When was the last time you played a AAA game on console that did not incorporate upscaling?

I don’t know. It’s what we got, I guess.

 

[Minsc]

When was the last time you played a AAA game on console that did not incorporate upscaling?

I also think AI upscaling can do far more than resolution. Like in terms of IQ, AA etc.

I have no doubt a 1440p image sent through AI could look more detailed than a native 4k image with inferior image smoothing methods.

I've seen time and time again, small details, edges, text etc, all can be near impossible to make out in a native res image, but when upscaled through AI everything can look vastly better.

 

[sendit]

Pure mathematical tasks. No complex reasoning or creative output.

AI is an iteration of that. A series of mathematical task/calculations and algorithms. If we relate this to Gen AI, it is pretty much a calculator with an artificial brain. The things it will eventually spew out will be imperceptible to real humans. I still don't see how this relates to your original argument, AI = Fad. We are just going down a rabbit hole on why you think AI sucks.

 

[Topher]

I also think AI upscaling can do far more than resolution. Like in terms of IQ, AA etc.

I have no doubt a 1440p image sent through AI could look more detailed than a native 4k image with inferior image smoothing methods.

I've seen time and time again, small details, edges, text etc, all can be near impossible to make out in a native res image, but when upscaled through AI everything can look vastly better.

Exactly. Resolution upscaling was really just the beginning. AI is going to have a positive influence on many aspects of graphical quality. Personally I can't wait to see what it can do.

 

[jroc74]

Calculators don’t try to pass themselves as something they’re not.

Oh so you were serious.

Someone tell Sony to stop working on Sophy ASAP.

Also...tell Rockstar to stop trying to improve on how NPC's behave in their worlds.

 

[sendit]

Exactly. Resolution upscaling was really just the beginning. AI is going to have a positive influence on many aspects of graphical quality. Personally I can't wait to see what it can do.

Yep, I don't even get excited about teraflop numbers. I'd rather see the next iteration of DLSS or similar.

 

[Topher]

Yep, I don't even get excited about teraflop numbers. I'd rather see the next iteration of DLSS or similar.

Yep. Folks gotta get used to the fact that GPUs are having to broaden their capabilties beyond rasterization. Isn't just about CUs and frequencies anymore.

 

[angrod14]

AI is an iteration of that. A series of mathematical task/calculations and algorithms. If we relate this to Gen AI, it is pretty much a calculator with an artificial brain. The things it will eventually spew out will be imperceptible to real humans. I still don't see how this relates to your original argument, AI = Fad. We are just going down a rabbit hole on why you think AI sucks.

Are you a replicant?

 

[MasterCornholio]

Of course the day after I say I think these specs are bullshit.......Tom Henderson confirms them.

OK go play ping pong with the communists.

 

[FireFly]

I don’t know. It’s what we got, I guess.

DLSS and XeSS both use machine learning to deliver much better image quality than FSR or other software upscaling techniques. So it's not just marketing BS but an actually quantifiable advantage. As PS5 games will use upscaling anyway, if the new technology is similar, it will just deliver a cleaner image with less artifacts.

 

[angrod14]

DLSS and XeSS both use machine learning to deliver much better image quality than FSR or other software upscaling techniques. So it's not just marketing BS but an actually quantifiable advantage. As PS5 games will use upscaling anyway, if the new technology is similar, it will just deliver a cleaner image with less artifacts.

The thing is, instead of pursuing the real deal we're comfortable settling with these techniques. They might be "better" than a lower native resolution but they still introduce artifacts and other shit.

 

[shamoomoo]

What's this shit? Since when have we been using TOPS as an a benchmark for gaming performance? Anyways, which TOPS are you talking about, INT4 or INT8? I'm pretty sure AMD will inflate these numbers as high as they can so they'll use INT4 numbers.

This is a slide from 2018 or whenever Turing was released:

So 228 TOPS INT8 and 455 TOPS INT4 for RTX2080 Ti.

Just like TFLOPS, implementation can affect how much performance a processor can achieve. If I'm not mistaken, Turing has higher ML performance than the comparable Ampere card for certain data types, does that imply Turing is a better architecture vs Ampere?

 

[FireFly]

The thing is, instead of pursuing the real deal we're comfortable settling with these techniques. They might be "better" than a lower native resolution but they still introduce artifacts and other shit. Nothing beats the real thing.

The "real thing" would be SSAA (running at 8K or 16K resolution then downsampling), not "native" 4K with TAA which either leads to vaseline-like smearing or ghosting. On PC, upscaling from half resolution with DLSS/XeSS generally gives similar quality overall to the "native" output, due to how bad TAA is.

And on top of that you get much better performance, so you can include more advanced lighting effects.

 

[DaGwaphics]

Pretty sure you use the shader units to calculate TF

PS5, for example.

2304 shader units times 2 operations per cycle times 2233mhz (freq) = 2304x2x2233=10,289,000/1,000,000=10.28TF

AMD Playstation 5 GPU Specs

AMD Oberon, 2233 MHz, 2304 Cores, 144 TMUs, 64 ROPs, 16384 MB GDDR6, 1750 MHz, 256 bit

www.techpowerup.com

If we do the same calculation for the RDNA 2.0 6800 XT....

4608 shader units times 2 operations per cycle times 2250mhz = 4608x2x2250=20,736,000/1,000,000=20.73 TF

Gaiff what is interesting about this is if we do the same calculation on 7700 XT then we do not get 35 TF

3456 shader units times 2 operations per cycle times 2544mhz = 3456x2x2544 = 17,584,128/1,000,000=17.58 TF

So the only variable that can be questioned here is the operations per cycle which relates to the dual-issue shaders you were talking about. What is differnt in RDNA 3.0 vs 2.0. Techpowerup calculates it differently as well for RDNA 3.0

AMD Radeon RX 7700 XT Specs

AMD Navi 32, 2544 MHz, 3456 Cores, 216 TMUs, 96 ROPs, 12288 MB GDDR6, 2250 MHz, 192 bit

www.techpowerup.com

It will be interesting to see how well the dual-issue shaders can be used in a console, so far it hasn't amounted to much in the PC space.